OpenAIAdapterOptions

Defined in: packages/realtime-openai/src/openai-adapter.types.ts:26

Options for constructing an OpenAIAdapter.

Properties

debug?

readonly optional debug?: boolean

Defined in: packages/realtime-openai/src/openai-adapter.types.ts:38

Enable verbose debug tracing (default false). Inbound provider events are logged as metadata only; transcripts, audio, identifiers, and raw payloads are never included.

---

greetingInterruption?

readonly optional greetingInterruption?: boolean | BrowserVadOptions

Defined in: packages/realtime-openai/src/openai-adapter.types.ts:58

Experimental local greeting-interruption configuration (default off). Omitted or false preserves the existing non-interruptible greeting gate. true selects shared defaults; an object selects validated browser-VAD settings. Enabled connections lazy-load the browser runtime. Local acceptance is playback-scoped: the first accepted local boundary stops greeting playback, then flushes one bounded pre-roll. Skip and normal completion discard the retained tail. Setup and worklet-command failure fall back to the non-interruptible gate.

---

mintToken

readonly mintToken: () => Promise<OpenAITokenGrant>

Defined in: packages/realtime-openai/src/openai-adapter.types.ts:32

Mints the ephemeral OpenAI realtime client secret this adapter connects with. Called once per connect(). The token handshake is assigned to the adapter — RealtimeSession never sees a credential.

Returns

Promise<OpenAITokenGrant>

---

serverVad?

readonly optional serverVad?: object

Defined in: packages/realtime-openai/src/openai-adapter.types.ts:79

OpenAI server-VAD tuning and barge-in policy.

The threshold defaults to 0.5 and must be a finite number in [0, 1]. Barge-in defaults to immediate, which preserves OpenAI-managed response creation and interruption. transcript-confirmed disables both provider flags so the adapter can interrupt on the first non-whitespace provisional transcription delta and create a response only after the same committed item has a non-whitespace completed transcript. Delta arrival latency is nondeterministic, and non-whitespace transcription does not prove that the sound came from a human speaker.

This configures OpenAI’s provider-side VAD only. LiveKit owns browser VAD, microphone gating, and turn boundaries through LiveKitAdapter’s separate vad option.

transcript-confirmed requires input transcription. Combining it with transcription: false throws a synchronous TypeError from the constructor before token minting, microphone capture, peer creation, or connection.

bargeInMode?

readonly optional bargeInMode?: "immediate" | "transcript-confirmed"

immediate (default) derives create_response: true and interrupt_response: true. transcript-confirmed derives both as false and requires input transcription.

threshold?

readonly optional threshold?: number

Finite OpenAI server-VAD activation threshold in [0, 1] (default 0.5).

---

synthesizeGreeting?

readonly optional synthesizeGreeting?: SynthesizeGreeting

Defined in: packages/realtime-openai/src/openai-adapter.types.ts:100

Fetches WAV bytes for the exact greeting text. Optional — without it the adapter cannot start greeting speech and returns not_started (no browser TTS fallback). Transport only; does not play audio.

---

transcription?

readonly optional transcription?: false | { model?: string; }

Defined in: packages/realtime-openai/src/openai-adapter.types.ts:47

Input-audio transcription configuration for the realtime session.

Omitted (the default): transcription is enabled with the built-in model. An object: transcription is enabled with model (or the built-in default when model is omitted). false: transcription is disabled and no transcription config is sent in session.update.

---

voice?

readonly optional voice?: OpenAIBuiltInVoice

Defined in: packages/realtime-openai/src/openai-adapter.types.ts:94

Built-in Realtime/TTS output voice. Defaults to marin. Applied to the initial session.update audio.output and passed to synthesizeGreeting.